卡在pass_level0

[1311523821]

error log | 日志或报错信息 | ログ

D:\ProgramData\anaconda3\envs\bridgedepth\lib\site-packages\torch\jit_trace.py:1307: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error:
Tensor-likes are not close!

Mismatched elements: 30631 / 30720 (99.7%)
Greatest absolute difference: 3.666351318359375 at index (0, 2, 11, 109) (up to 1e-05 allowed)
Greatest relative difference: 23.89518081229656 at index (0, 11, 12, 12) (up to 1e-05 allowed)
_check_trace(
pnnxparam = ./onnx/bridge_rvc_pretrain_model_ncnn.pnnx.param
pnnxbin = ./onnx/bridge_rvc_pretrain_model_ncnn.pnnx.bin
pnnxpy = ./onnx/bridge_rvc_pretrain_model_ncnn_pnnx.py
pnnxonnx = ./onnx/bridge_rvc_pretrain_model_ncnn.pnnx.onnx
ncnnparam = ./onnx/bridge_rvc_pretrain_model_ncnn.ncnn.param
ncnnbin = ./onnx/bridge_rvc_pretrain_model_ncnn.ncnn.bin
ncnnpy = ./onnx/bridge_rvc_pretrain_model_ncnn_ncnn.py
fp16 = 0
optlevel = 0
device = cpu
inputshape = [1,3,96,160]f32,[1,3,96,160]f32
inputshape2 =
customop =
moduleop =
############# pass_level0
inline module = bridgedepth.bridgedepth.BridgeDepth
inline module = bridgedepth.monocular.depth_anything.DPTHead
inline module = bridgedepth.monocular.depth_anything.FeatureFusionBlock
inline module = bridgedepth.monocular.depth_anything.ResidualConvUnit
inline module = bridgedepth.monocular.dinov2.layers.attention.MemEffAttention
inline module = bridgedepth.monocular.dinov2.layers.block.NestedTensorBlock
inline module = bridgedepth.monocular.dinov2.layers.layer_scale.LayerScale
inline module = bridgedepth.monocular.dinov2.layers.mlp.Mlp
inline module = bridgedepth.monocular.dinov2.layers.patch_embed.PatchEmbed
inline module = torch.nn.modules.linear.Identity
inline module = bridgedepth.bridgedepth.BridgeDepth
inline module = bridgedepth.monocular.depth_anything.DPTHead
inline module = bridgedepth.monocular.depth_anything.FeatureFusionBlock
inline module = bridgedepth.monocular.depth_anything.ResidualConvUnit
inline module = bridgedepth.monocular.dinov2.layers.attention.MemEffAttention
inline module = bridgedepth.monocular.dinov2.layers.block.NestedTensorBlock
inline module = bridgedepth.monocular.dinov2.layers.layer_scale.LayerScale
inline module = bridgedepth.monocular.dinov2.layers.attention.MemEffAttention
inline module = bridgedepth.monocular.dinov2.layers.block.NestedTensorBlock
inline module = bridgedepth.monocular.dinov2.layers.layer_scale.LayerScale
inline module = bridgedepth.monocular.dinov2.layers.layer_scale.LayerScale
inline module = bridgedepth.monocular.dinov2.layers.mlp.Mlp
inline module = bridgedepth.monocular.dinov2.layers.patch_embed.PatchEmbed
inline module = torch.nn.modules.linear.Identity

---

context | 编译/运行环境 | バックグラウンド

ncnn 1.0.20250503
pnnx 20251016

卡在pass_level0一般是什么原因

[dosubot]

Hi @1311523821! I'm Dosu and I’m helping the ncnn team.

“卡在pass_level0”通常说明 pnnx 在模型转换的早期阶段（模块内联、输入结构展开、shape 推理等）遇到了无法处理的情况，导致流程无法继续。常见原因包括：模型里有复杂或自定义模块（比如 bridgedepth、dinov2 等），这些模块结构动态、包含特殊操作或 PyTorch 高级用法，pnnx 可能无法正确内联或推理其 shape，进而卡住或报错。还有可能是输入 shape/dtype 不匹配、模型参数有零维度、chunk 操作维度不对、或者模型里有未支持的操作等[pnnx README]。

建议排查：

• 检查模型输入 shape 和 dtype，确保没有零维度或异常值。
• 尝试用 moduleop 参数让 pnnx 保留部分复杂模块为整体算子（比如 moduleop=bridgedepth.bridgedepth.BridgeDepth,bridgedepth.monocular.dinov2.layers.block.NestedTensorBlock），避免内联失败。
• 检查模型是否包含未支持的 PyTorch 操作或自定义 C++/CUDA op。
• 尝试用更简单的模型或逐步去掉模块定位具体卡住的地方。

如果有更详细的报错信息或日志，可以进一步定位是哪一个模块或操作导致卡住。

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}